Topos Bio, an AI-powered biotechnology company, has introduced two new AI foundation models that could help predict how new proteins can be used to treat previously “undruggable” diseases.
The first AI model, Topos-2, is built to predict how intrinsically disordered proteins (IDPs), which are dynamic in nature, move and change shapes. The second AI model is Topos-Bind, a first-of-its-kind model designed to predict how these dynamic proteins interact with potential drug molecules, said Topos Bio CEO Ryan Zarcone and CTO Amir Khosrowshahi, in an interview with GamesBeat.
Together, these models represent a significant advancement in AI-driven drug discovery, addressing a class of disease-causing proteins historically difficult to study and target. That includes Parkinson’s Disease.
Topos Bio’s mission is to build foundation models to drug historically “undruggable” targets, disease-causing proteins that resist traditional drug therapies and tools.
“If you kind of think about some of the the most impactful diseases today facing society that still don’t have drugs, neurodegenerative diseases are one of those top categories,” Zarcone said. “There are other areas like very aggressive forms of cancer, autoimmune disease, but we will just take neuro for now and think about that.”
“So, if you ask the question, ‘Why is it that there are still no drugs for these diseases? Our perspective really comes down to dynamics. I’ll unpack it. If you think about really what is the cause of any disease in humans, it’s basically the breakdown of some machinery inside of a cell in neurogenesis like Alzheimer’s and Parkinson’s.”
How it works
Our brain cells are neurons. In diseases like Parkinson’s, those proteins in the neurons are very dynamic. They’re highly flexible, and they are constantly changing shape.
“The challenge has been that traditional drug discovery and drug design has been really focused on static proteins,” he said.
Perhaps two thirds of the proteins in humans are static and structured. But a third or so are really dynamic — disordered in formal medical talk.
“These are the exact proteins that are implicated in these diseases, like Parkinson’s disease. We started Topos Bio because we saw this and saw that there really aren’t any good tools for modeling these dynamic proteins or for designing drugs for them,” Zarcone said. “These two problems are related.”
The team looked at the latest AI models for drug design, but they worked best for structured or static proteins, not for dynamic proteins. That’s where Topos Bio started.
“We’ve basically built a full stack drug design engine for these dynamic proteins,” Zarcone said. “We create in-house large quantities of both simulation data and experimental data, and then we train foundation models on top of that data, and then we use those foundation models to design drugs for proteins which were previously undrugged, like Parkinson’s disease, for example.”
Zarcone said the model release that we just announced is basically trying to attack these two problems. Topos-2 is foicused on how to model these really dynamic, flexible proteins across the human proteome for all the proteins that are these dynamic proteins.
And Topos Bind, the other foundation model, focuses on designing drugs that can work on the dynamic proteins.
“To our knowledge, Topos Bind is the first model that’s capable of doing this, which is really exciting,” Zarcone said. “We see this as the first steps on a much longer journey to come up with useful simulations of biology and to use those to design drugs.”
The company is currently figuring out how it makes its tech available as a product.
“This wouldn’t have really been possible even like a few years ago,” Zarcone said. “There can be accelerated processes for getting into the clinic because to to encourage people to work on some of these rare genetic diseases.”
These targets, comprising close to two-thirds of all proteins in humans, are major drivers of the most aggressive cancers and neurodegenerative diseases including Alzheimer’s, Parkinson’s, ALS, and Huntington’s. They represent a combined commercial opportunity of more than $100 billion a year, and until now, the industry has had no rational way to design drugs against them. That’s why only a handful of drugs have been produced that halt disease progression at all.
“The hardest problems in biology aren’t hidden because we’ve lacked computing power, they’re hidden because we’ve lacked the right data and models,” said Ryan Zarcone, CEO and co-founder of Topos Bio. “We’ve built Topos around a simple premise: if we can understand how these proteins move and interact, we can design medicines for diseases that have long remained beyond the reach of modern drug discovery, and ultimately change the trajectory of patients’ lives.”
Topos-2 builds on Topos-1, its predecessor which generates conformational ensembles for disordered proteins and regions. Topos-2 extends this capability to proteins that have both disordered and ordered regions, a closer match to how these proteins actually work in the body.
On PeptoneBench, an external experimental benchmark measuring performance on these dynamic proteins, Topos-2 ranked first among over 10 evaluated models on agreement with experimental data, including NMR chemical shifts and SAXS.
It outperformed other frontier models such as Microsoft Research’s BioEmu, Google DeepMind’s AlphaFold-2, and Boltz’s Boltz-2. Early testing put it over four times closer to experimental ground-truth than the next-best model and showed that Topos-Bind was the only model able to predict how the shape of amyloid-beta, an Alzheimer’s target, changes in response to drugs.
Where Topos-2 models these dynamic proteins on their own, Topos-Bind predicts how they interact with drugs, giving deeper insight into new ways to target disease-linked proteins. No other existing model does this today. To train it, Topos Bio assembled the largest experimental datasets of small-molecule binding measurements against IDPs – at least two orders of magnitude larger than previously reported datasets – along with the largest physics-based corpus of IDR-ligand systems.
“Most AI protein models predict one best structure per sequence, but disordered proteins don’t have one shape, they exist as a shifting ensemble, so a single-structure model is answering the wrong question,” said Khosrowshahi, CTO. “Topos Bio’s models generate a full ensemble of conformations a protein can actually adopt. From there, the models learn how a drug acts across that shifting landscape, not just where it binds in one snapshot, but how it changes the protein’s behavior.”
Topos-2 and Topos-Bind are currently being applied across Topos Bio’s internal discovery programs in neurodegeneration and oncology. That real-world application is already reshaping how the team thinks about the diseases they’re targeting.
“Many of the proteins most relevant to neurodegenerative disease are also among the hardest to study because they do not remain in one shape,” said Steven Finkbeiner, a medical doctor and director of the Center for Systems and Therapeutics at Gladstone Institutes and a scientific collaborator with Topos Bio, in a statement. “For years, the field has lacked sufficient experimental data to connect the motion of these proteins to their molecular interactions. Combining large-scale data generation with AI models designed for dynamic proteins could help address this gap and advance our understanding of disease biology and the search for new therapeutics.”
Topos Bio purposely chose to start with intrinsically disordered proteins, the extreme version of this problem, because solving it first forces the right approach from day one. That same framework is built to extend to the broader universe of dynamic, context-dependent drug targets that conventional discovery has never been able to reach.
Origins
The founders are neuroscientists who were in the same circles, though at different times in their careers.
“Amir and I are neuroscientists by training, and we were in the same PhD lab — just at separate times. We had the same PhD adviser in neuroscience. We were in a computational and theoretical neuroscience lab, so we had a very long-standing interest in neuro,” said Zarcone.
Over time, they learned that neurodegenerative diseases and other conditions involving dynamic, disordered proteins have been poorly modeled by traditional drug-discovery tools.
The San Francisco company started in January 2025 after closing a $10.5 million seed round led by Threshold Ventures and Boldstart, with participation from Neo and several angel investors such as former Intel exec Naveen Rao.
The team has 16 people, roughly half of them trained in computational or AI work and the other half trained in biological labs. It also has a team of six in China. That combination of disciplines is part of the interesting technical challenge facing Topos Bio.
“The wet lab scientists are doing two things: they’re generating experimental data to train our foundation models, and then they’re actually taking predictions from the models and testing them in the lab,” Zarcone said. “So we have these models that generate candidates for drugs, and then we synthesize those, and then we test them in a variety of different experiments, to see if they actually were functional and actually slowed disease progression, or markers of disease progression in cells.”
The platform combines protein simulations, experimental data, foundation models, and cellular testing. Simulations of up to 40,000 atoms run via distributed cloud computing rather than a dedicated supercomputer. Testing includes neurons derived from patient blood or skin cells through induced pluripotent stem-cell methods, including samples from the Gladstone Institute at UCSF, said Khosrowshahi.
The company gets access to brain cells using a partnership with the Gladstone Institute at UCSF, where Topos Bio works with neuroscientist Steve Fink. They have a bio bank of samples from patients with various neurodegenerative mutations. So they have a collection of samples from patients with Parkinson’s mutations.
“Those are the cells we are working with,” Zarcone said. “For the data collection efforts that we have, we’re planning to scale up anywhere between 10 times to 100 times what we’ve collected thus far.”
The future
Topos Bio is in evaluation discussions with several large pharmaceutical companies about embedding the models into drug-design workflows; no formal partnerships have been announced yet, but the founders say they are in progress.
The company plans to scale data generation by approximately 10 to 100 times over the next six months and expects potential FDA submissions within one to two years. On those submissions, it will likely work with an external biotech or pharmaceutical partner.
“We’ve been talking to a number of big pharmas, who are interested in these areas to work toward taking these models and actually just embedding them into their workflows or in different groups across a bunch of therapeutic areas,” Zarcone said.